README.md

GradPath Readme

This repository gives details for the paper "Generating and Evaluating Plausible Explanations for Knowledge Graph Completion", ACL 2024.

Datasets

We use the following two datasets:

• Kinship from Hinton 1990
• CoDEx from Safavi & Koutra EMNLP 2020

Explanation Methods

Baseline

The baseline explanation method that we use is Gradient Rollback. Please follow the instructions in this repository to generate explanation with this method.

GradPath

Follow step 1 in Gradient Rollback to create a model and corresponding influence map. GradPath provides an alternative method to compute step 2 of Gradient Rollback.

GradPath Explanations

Hyperparameter Settings

To train the Kinship model following step 1 in Gradient Rollback:

--epochs 100 \
--latent_expert_embedding_dim 10 \
--batch_size 1 \
--num_negative 13 \
--top_k 1 \
--optimizer 'adam' \
--learning_rate 0.001 \
--train_with_softmax True \
--expert 'ComplEx' \

To train the CoDEx model following step 1 in Gradient Rollback:

--epochs 500 \
--latent_expert_embedding_dim 1024 \
--batch_size 1 \
--num_negative 20 \
--top_k 1 \
--optimizer 'adam' \
--learning_rate 0.3 \
--train_with_softmax True \
--expert 'DistMult' \

Explanation files

Instead of generating gradient rollback explanations, we want to generate GradPath explanations. This can be done by following step 2 in Gradient Rollback, Please replace the for loop with the below code and make sure that gradpath_explanations.py is available.

    prediction_triple = predictions[test_triple_idx][0]
    prediction_str = ' '.join(prediction_triple[0:3])

    current_prediction = {}
    current_prediction["correct"] = one_if_top1_correct[test_triple_idx]
    current_prediction["probability"] = predictions[test_triple_idx][0][3]

    topN_explanations = {}
    paths_all = get_all_paths(lookup_table, prediction_triple[0], prediction_triple[2], nodelimit)
    for ell in range(1,nodelimit):
        explanations = get_explanations(model_holder, influence_map, prediction_triple, paths_all, length=ell)
        if explanations is not None and len(explanations)>TopN:
            tem = sorted(explanations.items(), key=lambda item: item[1],reverse=True)
            for tem_i in range(TopN):
                k,v = tem[tem_i]
                topN_explanations[k] = v
        elif explanations is not None and len(explanations)>0:
            topN_explanations.update(explanations)
    current_prediction["explanations"] = topN_explanations

    explanation_dictionary [prediction_str] = current_prediction